A man named Amir Adivi called his insurance company's phone line 24 times. On seven of those calls, he says, he waited longer than the law allows for a human to pick up. On four of them, he waited over 20 minutes. His lawyer used those calls to try to open a class action against the insurer, AIG Israel, on behalf of every customer who had ever waited too long. In August 2026, that case ended, not with a ruling on the merits, but with a judge's finding that his lawyer's court filing rested on legal precedents that were never real.
The claim itself
Israeli law requires insurers to provide a human phone response within six minutes for certain kinds of calls. Adivi argued that AIG routinely blew past that limit, sometimes by half an hour, and that this added up to a pattern affecting the company's entire customer base. AIG's answer was that its performance is measured as a yearly average across all calls, not call by call, and that regulators already track this exact number.
That argument, about how the law measures compliance, is the kind of dispute courts handle every day. It's what happened next that made this case notable.
The citations that weren't there
At a pre-trial hearing, AIG's lawyer told the court that several of the precedents cited in Adivi's lawyer's filing didn't check out. The judge, Hadas Ovadia of the Tel Aviv District Court, looked into it herself. She found five citations to rulings that, as she put it, she could not confirm existed at all. Among them:
- A ruling attributed to an appeal between Phoenix Insurance and a customer, said to establish a legal principle.
- A class action against a telecoms company, cited without any way to locate it.
- A ruling supposedly holding that wasted time and systematic inconvenience can count as compensable harm.
- A ruling supposedly recognizing non-financial consumer harm even when each individual's loss is small.
- A ruling supposedly finding that a pattern of complaints can itself point to a systemic failure.
Each one, the judge wrote, appeared designed to support exactly the argument the filing was making. None of them could be found. She also noted a smaller tell: parts of the filing read as though a single sentence had been split across four separate numbered points, the kind of formatting a person editing their own writing rarely produces.
What the judge actually decided about AI
The most interesting part of the ruling isn't the discovery itself. It's the line the judge drew afterward. She was explicit that she was not ruling against the use of artificial intelligence in legal work as such. Her comparison: using AI is a shift some would place alongside the industrial revolution, and mistakes can happen "in operating any machine, including this one." What drew her criticism was narrower, a use of the tool that was blind, systematic, and never checked against a real source.
In practice, that distinction shaped her decision. She considered ordering the lawyer to personally pay costs, a more serious step than making his client pay, and chose not to. Instead she issued what amounts to a formal warning, paired with a costs order against his client of 20,000 shekels, payable to the insurance company. The underlying class action was withdrawn.
Part of a pattern, not a one-off
The ruling itself points to three earlier Israeli cases where courts had already confronted the same problem: fabricated citations appearing in filings before a religious appeals court, before a court reviewing a municipal decision, and before the High Court of Justice in a case about animal welfare policy. Judge Ovadia's ruling adds a fourth. Read together, they suggest a recurring pattern moving through the Israeli court system rather than an isolated mistake, and a judiciary that is starting to build a consistent response to it.
Why it isn't just a story about lawyers
The underlying failure here isn't specific to courtrooms. A tool confidently produced citations, names, and legal reasoning that sounded exactly right and were entirely invented, and a professional trained to check sources filed them anyway. That is the same failure mode anyone risks when they take a confident answer from an AI tool at face value instead of verifying it against something real. The court's response was not to ban the tool. It was to insist that a human stay responsible for what it produces.